Will transformer give you answers? An effective way to conduct multilingual real-world ConvQA tasks with transformer

Abstract

In real-world applications, the Conversational Question Answering (ConvQA) task still faces the challenge of how a machine can answer questions in the absence of explicit knowledge and how to make machines efficiently select and encode both the current question and historical contexts in multiple rounds of conversation. This paper presents a transformer-based retrieval-reading system with customized modules to investigate the possibility of using background knowledge to answer questions and explore a few directions to leverage historical contextual information in real-world multilingual ConvQA scenarios. Overall, our experimental results show that the machine can create significantly better answers when background knowledge and refinement of historical contexts are taken into account.